ci: stop the test cache self-invalidating, and drop duplicate PR runs

[ackness]

CI-only change. No version bump, no release — docs/CHANGELOG.md and the version badges are untouched.

The problem

turbo run test never cached. Two consecutive runs with no edits in between:

Cached:    2 cached, 43 total
  Time:    49s

test listed $TURBO_ROOT$/plugins/** as an input, which pulled each plugin's run artefacts into the hash. Diffing two run summaries:

### DIFF in inputs
  ../story-guard/node_modules/.vite/vitest/…/results.json:  94b73c9e → 8823c1d7
  ../char-creator/.turbo/turbo-test.log:                    0d437d09 → 4866ff05
  …41 entries, every plugin package
environmentVariables:        identical
hashOfExternalDependencies:  identical
dependencies:                identical

Those are vitest's result cache and turbo's own task log. Running the suite rewrites them, so the next run hashes differently and all 43 packages miss — running tests was the thing invalidating the test cache.

They are gitignored, but explicit input globs don't consult gitignore the way $TURBO_DEFAULT$ does. That's why git status stayed clean while the hash moved, and why this was invisible.

The fix

Negative globs (!.../node_modules/**) don't work here — inputs resolve per package and these are cross-package paths; verified, still 21/43. So the directory glob is replaced with a whitelist of the source shapes a plugin actually has: markdown, entry .js, package.json, and the eight content directories (hooks lib runtimes schemas server tests tools ui). Artefact directories can no longer match.

Verification

pnpm test ×2, no edits between:   2/43 cached, 49s  →  43/43 cached, 0.3s (FULL TURBO)
edit plugins/narrator/PLUGIN.md:  2/43 cached          ← correctly invalidated
revert it:                        43/43 cached
turbo run lint  ×2:               21/21 FULL TURBO
turbo run build ×2:               4/4   FULL TURBO

Dependency semantics are unchanged — editing any plugin source still invalidates every dependent package's test. The full suite passed on the run that actually executed (43 successful, 0 failed). actionlint and prettier pass.

Also here

• CI caches .turbo/cache. It ran cold every time, so none of the above would have reached CI otherwise.
• CI push trigger scoped to main. A *-dev branch with an open PR ran the same commit twice — the PR ref and the branch ref hash to different concurrency groups, so neither cancelled the other. Seen on the v0.0.21 release: runs 30162296595 (push) and 30162298111 (pull_request), same commit, ~5.5 min each. The PR run covers dev branches; the push trigger stays as the post-merge guard on main.
• macOS release job gets the Electron binary cache the Windows job already had.

Expected effect

Cold on the first run either way. From the second run on, a change touching one package replays the rest from cache instead of re-running them — the 155s Test step and 134s Build step should collapse to whatever actually changed.

---

中文摘要：纯 CI 改动，不发版。test 的 inputs 把 plugins/** 整个目录纳入哈希，连带 vitest 的 results.json 和 turbo 自己写的 turbo-test.log 也算了进去 —— 跑测试这个动作本身就让测试缓存失效，所以一直是 2/43。改成源文件白名单后 43/43 全命中（49s → 0.3s），且改插件源码仍能正确失效。另外给 CI 加了 .turbo/cache 缓存、把 push 触发收到只剩 main（消除 PR 双跑），并给 release 的 macOS job 补上 Windows job 早就有的 Electron 缓存。